Skip to navigation

Release Notes

October 29, 2026

v0.0.140

NemoClaw v0.0.140 improves sandbox recovery, onboarding and inference checks, Hermes credentials, and managed Deep Agents Code. It also makes backup, uninstall, and installer failures easier to recover from.

  • The managed Deep Agents Code image now uses version 0.1.87. Its initial model request retains search_tools when an upstream request projection omits that tool, so managed MCP search remains available. Related changes: PR #13878 and PR #13766.
  • Hermes images now include Switchyard 0.3.0 for qualification. The package remains disabled by default and does not enable a new route. Related change: PR #13879.
  • Hermes rebuilds preserve managed OpenRouter credential references while removing raw or unrecognized references from restored native state. Hermes also accepts the managed NVIDIA key template when its native credential slots agree. Related changes: PR #13771 and PR #13785.
  • Hermes restores the default profile’s SOUL.md and skills index. When neither Tavily nor the Nous-managed web gateway is configured, Hermes does not offer web_search or web_extract. Related changes: PR #12386 and PR #13620.
  • Hermes configuration export preserves a disabled dashboard and omits its port, which v1 rejects when the dashboard is disabled. It validates retained dashboard URLs against the allocated port. Related change: PR #12813.
  • A model-only change on a recorded DNS-backed HTTPS adapter route now verifies the selected model from the sandbox before it updates agent configuration or registry state. If verification fails, NemoClaw attempts to restore the previous OpenShell model selection and reports the result. Related change: PR #13237.
  • Local inference onboarding uses NEMOCLAW_LOCAL_INFERENCE_TIMEOUT to bound each applicable in-sandbox request, with a 60-second minimum. Managed llama.cpp fingerprint requests have a longer bounded deadline than operator-managed servers. Related changes: PR #13670 and PR #13561.
  • The Ollama model menu preserves a reported 0 free-GPU-memory reading instead of treating it as unknown. Managed vLLM publishes provider keys atomically, so a failed update does not leave a partial key set. Related changes: PR #13748 and PR #13137.
  • NVIDIA endpoint validation during onboarding or resume uses the native endpoint when saved metadata still names inference.local or has no endpoint. A provider with only a binding and no ownership receipt is treated as detached. Related changes: PR #13571 and PR #13649.
  • If onboarding finds a native provider without a NemoClaw ownership receipt, its error gives an openshell provider delete command with the gateway flag when needed. Check that no sandbox on that gateway uses the provider before running it. Related change: PR #13669.
  • Onboarding preserves custom endpoint URL, API family, and credential variable name at the configuration review checkpoint. It rejects unsupported Deep Agents Code auto-approval choices before startup and rejects invalid NEMOCLAW_CPU or NEMOCLAW_RAM percentages. Related changes: PR #13779, PR #13567, and PR #12422.
  • Sandbox status, recover, and connect --probe-only identify the target gateway. Status stays responsive during rebuild and marks inference as not probed when a safe read-only lifecycle snapshot is available. Related changes: PR #13134 and PR #13009.
  • destroy sends cleanup to the sandbox’s recorded gateway without changing the active gateway selection. Sandboxes on other gateways do not block final cleanup of the target gateway. Related changes: PR #13012 and PR #13819.
  • backup-all retries native-home capture when a stopped sandbox is still booting and the capture reports process churn. Other capture failures are not retried. Snapshot sanitization also accepts backslashes in recorded paths. Related changes: PR #12891 and PR #12381.
  • Deep Agents Code image builds now keep their npm cache outside /sandbox, so a restored agent home’s .npm state does not collide with the build cache. Related change: PR #13263.
  • Uninstall removes a dead gateway service on headless Linux and skips an absent default gateway during a whole-host sweep. It reports when the container engine does not respond. Related changes: PR #13630, PR #13446, and PR #13645.
  • The installer distinguishes a timed-out release lookup, an unavailable requested reference, and a Git verification failure. A managed install or upgrade can remove the existing CLI source checkout before a failed lookup; fix the lookup issue and rerun the installer to restore the command. Related change: PR #11531.
  • Deep Agents Code edit approval cards show the full requested path for a nonempty edit_file path. NemoClaw writes POLICY.md only in OpenClaw sandboxes. Related changes: PR #13591 and PR #13713.
  • A piped nemoclaw <name> agent message must finish within its host-side input deadline and be no more than 4 MiB. A timeout, oversized message, or read failure exits without starting an agent turn. Related change: PR #13646.
  • Ordinary installer and supported CLI operations attempt to send telemetry about completed operations to NVIDIA’s fixed TEST receiver by default. Production delivery is not enabled. Set NEMOCLAW_DISABLE_TELEMETRY=1 in each shell before installation or CLI use to disable it. Related change: PR #13772.
  • CI operations now attempt to send telemetry to the TEST receiver with a source-run label. If GitHub Actions environment values pass format checks, the label includes their reported run ID and attempt. Other eligible CI runs use a generic label. Opt-out, malformed explicit QA labels, and test-process suppression still apply. Related change: PR #13770.